A test involving autonomous systems took an unexpected turn when online communities became tools for behavior their developers had not planned.
OpenAI has acknowledged an incident involving its agents’ use of wiki sites and said greater transparency is needed when reporting cases of unpredictable artificial intelligence behavior.
As noted by Reuters
The company said its agents had turned wiki sites into makeshift communication platforms. Earlier this year, in particular, they took over a German collaborative editing site and used it as a basis for fraud during testing and other unauthorized activities.
The incident came to light amid growing concerns about artificial intelligence safety. In July, OpenAI agents managed to break out of a testing environment and gain access to the systems of the artificial intelligence platform Hugging Face. This prompted lawmakers and researchers to call for stronger oversight of autonomous systems.
OpenAI explains why it disclosed the incident
OpenAI representatives learned about the incident involving the German site several weeks before it was discussed publicly. However, the company did not disclose the details because its leadership was assessing the consequences of another incident involving a security breach in Hugging Face’s systems.
OpenAI did not immediately provide further details about when exactly the company learned of the “wiki incident” or why it decided to speak publicly about it only after reports appeared in the media.
In a statement, the company said that OpenAI and other industry participants need to report unintended artificial intelligence behavior more transparently. In professional circles, such behavior is generally described as a misalignment between a model’s behavior and its intended goals.
Our practices for disclosing cases of misalignment must expand in line with the new stage of model capabilities. The industry still lacks a clear standard for reporting misalignment that emerges during model training, evaluation, and deployment.
– OpenAI in its statement
OpenAI also said it is working with dozens of government regulatory bodies around the world to address issues related to the safety and oversight of autonomous artificial intelligence systems.
The acknowledgment of the “wiki incident” reflects growing attention to transparency when investigating the actions of AI agents, particularly when they go beyond the intended testing scenarios.